loopcat

Build in public!

[loopcat @loopcat101]

I think this is fake. Jev can help you choose but cannot generate the visualization after try-on. If you use a real-time video generation model, then I will say it is true.

[Nailthy Tang @nailthy62]

jev is insane 🫣

it makes realtime virtual try-on hauls possible.
built this experiment for Drape with @typesafeai

> i talk
> jev reads transcript + what i'm wearing
> picks from my closet
> changes my outfit in realtime

cost: $0.0011 per decision
time: ~620ms per decision

imagine getting ready like this:

[loopcat @loopcat101]

@passluo 平头哥

[loopcat @loopcat101]

Jev非���适合有限动作空间的应用，与以前的分类器不一样的是，这是一个openend分类器，input是freeform，类别也是freeform。

[Wey Gu 古思为 @wey_gu]

这两天在着手做关于 Jev 的工作时候，还是会感慨它的意义非常非常大

一方面，它的输出范式引导我们把决策与生成分离，在很多任务 harness 设计中，即便是用��一个 LLM 做这样的抽象都会有更好的效果，后边我们会分享我们的一些研究

另一方面，就是 Jev 找到了一种 LLM 程度泛化、上下文学习能力的协处理的模型的训练方法，打开了各种“子神经系统”的新智能系统设计范式，像我们让被习得的能力 offload 到反射、或者小脑/植物神经等生物子系统那样

比如不同的几个 Jev-like 模型（至少一个要带上视觉） 和一个 frontier LLM 的特制 harness ，我认为已经可能拼出来非常通用智能的机器人了，我不熟悉机器人的各个系统的接口，但是他们的反馈决策机制用，我猜测用通用的 Jev-lile 模型，是可能存在很可观的低垂果��达到从前需要非常昂贵训练的效果的？

这里，Laya 这一类的工作很棒，也让我特别兴奋，不过，似乎 Jev-like 的模型中，拥有更高泛化能力的（预训练更大的模型），延迟也大一些的模型也会有很多超出简单环境上快速反射复杂任务的场景，这样的工作可能不是简单拿大一些的开放权重 LLM 做 Jev-like 后训练能做到的？

另外，我认为从 LLM/Harness 大脑去低延迟协调多个子 Jev-like 模型组成的子神经系统是很有意义的研究，不知道这样的工作用多个模型联合组成是不是会比 harness 可编程粘合的方式个更有优势（不恰当类比 e2e 的 voice model vs 组合而来的 voice agent framewoek）

感慨，生物的高级神经系统演化是从协处理低级神经系统到有更复杂的大脑系统的，我们在人工智能领域竟然是先有大脑级别的模型，而后有通用的子神经系统的，非常有趣

[loopcat @loopcat101]

@criscxuan 看起来太gpt风格了

[loopcat @loopcat101]

@0xSero 5500 is coming.

[loopcat @loopcat101]

@kunchenguid The only problem with SuperGrok Heavy is that Grok still has a large gap compared to GPT and Claude.

[loopcat @loopcat101]

@FeitengLi 模型量化是最有效的加速手段。在cv小模型时代就是如此了，llm时代更加凸显出了量化节省显存的价值。

[loopcat @loopcat101]

Congratulations!

[Daphne Cornelisse @daphnesolves]

A little late, but happy to share that spiced self-play was accepted to the Conference on Robot Learning (CoRL)! 🎉

[loopcat @loopcat101]

@fkysly @realWeZZard 训练用Mac不行的，最好租卡

[loopcat @loopcat101]

@MiaAI_lab 552B Parameters🤣 We need 4x DGX Spark.

[loopcat @loopcat101]

DeepSeek-V4.1-Flash: 552B MOE
1. Asymmetric encoder-decoder structure: Input 8B activate, Output 16B activate
2. Reduction of KV Cache Per Token

[loopcat @loopcat101]

@gyrojeff_ scaning a fast moving train and get clear image?

[loopcat @loopcat101]

@Lingxiao234 Impressive! Going to try and reproduce this.

[loopcat @loopcat101]

DriveZero: an end-to-end framework beyond human demonstrations!

1. We build DriveRL: a policy model trained from scratch with closed-loop reinforced learning.
2. We train a visual foundation model DriveVFM by distilling DINO, SAM, SigLip2 and DepthAnything.
3. Finally, we introduce DriveZero, with DriveVFM as backbone, DriveRL as teacher.

[Haochen Tian @HCTian713]

1/n Intoducing #DriveZero 0️⃣ End-to-End Driving Beyond Human Demonstrations

https://t.co/PRNHpRLIVQ

🏆 A visual e2e driving policy that learns without human traj supervision, yet surpasses human driver on NAVSIM and sets SOTA on NAVSIMv2 & HUGSIM

🌍 Real-World Deployment ⬇️

[loopcat @loopcat101]

GPT-6 Astra’s spatial reasoning and robotics demos have taken over my feed. Beyond the impressive results, they’ve made me think about what general-purpose agents bring to Physical AI.
Their biggest advantage, in my view: turning feedback into better ways of solving unfamiliar tasks.
VLA/WAM policies can use feedback and recover from failures too. What excites me about general-purpose agents is the breadth of their problem-solving abilities: diagnosing failures, forming hypotheses, changing strategies, rewriting control code, and testing whether those changes actually work.
I saw a small example when I used Codex to control a simulated robot arm from images. It missed a grasp, repositioned, then tipped a bowl during release. On the next attempt, it adjusted its approach and completed the stack. No weight updates—but failure changed its subsequent behavior.
With informative feedback, recoverable failures, and enough time to experiment, I believe agents can work their way through many manipulation tasks they initially cannot solve.
Better spatial intelligence improves both first-attempt success and the quality of corrections. The agent loop lets those attempts build on accumulated experience.
My bet: the future of Physical AI will have two systems. A general-purpose agent handles goals, reasoning, and adaptation, while tools handle perception and execution. VLAs, WAMs, 3D perception systems, and controllers all become tools the agent can use.
Those tools are the agent’s hands and feet in the physical world.

[loopcat @loopcat101]

Maybe they have not used Linux before 😂

[Chaofan Shou @shoucccc]

i interviewed a few new grads asking for $150k+ salary and had them to ssh into a server

three sent me their ssh private keys, two never heard of it, one tried to rdp the host, one gave me their computer's teamviewer token. some of them are phd btw.

[loopcat @loopcat101]

If you choose max effort, you will get 1.5x or more usage for opus, and 3.5x for fable.

[loopcat @loopcat101]

OMG! Fable 5.1 is coming!

[Claude @claudeai]

We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.

They're the world’s most advanced models for coding and knowledge work.

[loopcat @loopcat101]

What is the price of dgx spark in your country?
In china, it’s about 3.3w RMB (about 4900 usd)